Objective Against the backdrop of high - density urban development, residents’ mental health problems have become increasingly severe. Access to urban green spaces is widely regarded as an important approach to improve residents’ mental health. Exploring the impact of green space characteristics on mental health benefits can provide a theoretical basis for urban green space planning and design from the perspective of healthy cities. This study aims to clarify the internal relationships between objective and subjective green space characteristics and different mental health benefits (emotional restoration, cognitive enhancement, and stress relief) through explainable machine - learning models.

Methods A mental health perception restoration experiment was carried out in two green spaces (Yanziji Park and Xiamafang Park) in Nanjing. Fifty - six participants engaged in two - hour free activities in the green spaces. During this period, GPS trajectories, objective green space characteristic data, subjective green space characteristic perception assessment data, and self - assessment data of mental health benefits were collected. Objective green space characteristics included the Normalized Difference Vegetation Index (NDVI), green view ratio, canopy density, actual noise dB (A), and spatial attractiveness, which were measured by remote sensing, semantic segmentation, and acoustic instruments. Subjective green space characteristics, such as perceived greenness, perceived noise, and perceived attractiveness, were evaluated by means of a 5 - point Likert scale questionnaire. Mental health benefits were divided into three types: emotional restoration, cognitive enhancement, and stress relief, and were assessed using the Restorative Outcomes Scale (ROS). To analyze and clarify the relationships between objective and subjective green space characteristics and different types of mental health benefits, the study adopted the Light Gradient Boosting Machine (LightGBM) model, combined with SHapley Additive exPlanations (SHAP) to measure and explain the importance of green space characteristics for mental health benefits. Based on the SHAP values, the non - linear relationships between them were further clarified.

Results Through the analysis of 3 types of mental health benefits and 5 models, the LightGBM model outperformed other algorithms (such as Random Forest and XGBoost) in terms of prediction accuracy (R²: 0.523 - 0.642), verifying its robustness in capturing complex feature interactions. The SHAP value analysis showed that subjective green space characteristics had a stronger relative impact on mental health outcomes than objective indicators. Among them, perceived attractiveness was the most important contributing factor, followed by perceived greenness and perceived noise. Notably, the positive impact of perceived greenness on mental health was greater than that of objective indicators such as the green view ratio and NDVI. In addition, in terms of noise, excessive actual noise could inhibit cognitive enhancement and stress relief. However, moderate perceived noise could promote emotional restoration and stress relief. For example, in the cognitive enhancement model, when the actual noise exceeded 53.88 decibels and in the stress - relief model, when it exceeded 52.73 decibels, negative effects would occur. While in the emotional - restoration model, when the perceived noise was within a certain range (less than 2.58 points), it was beneficial for emotional restoration.

Conclusion The results of this study provide empirical evidence for the internal relationship between urban green spaces and residents’ mental health. Firstly, this study constructed an index system covering both objective and subjective characteristics. By combining field measurements, questionnaire surveys, and advanced machine - learning algorithms, it explored the impact of green space characteristics on emotional restoration, cognitive enhancement, and stress relief. Secondly, subjective green space characteristics play a prominent role in influencing mental health benefits. The combined influence of perceived attractiveness and perceived greenness is the most significant. The results of non - linear regression show that actual noise has an inhibitory effect on cognitive enhancement and stress relief, while moderate perceived noise can promote emotional restoration and stress relief. Finally, this study provides a direction for further exploring the deep - level association mechanism between green spaces and mental health, and also offers data support for urban green space planning and design aimed at promoting residents’ mental health.